Community Participation in Decision-Making
Evidence from an experiment in providing safe drinking water in
Bangladesh
Malgosia Madajewicz∗1, Anna Tompsett2 and Ahasan Habib3
1NASA Goddard Institute for Space Studies, Columbia University, New York2School of International and Public Affairs, Columbia University, New York
3NGO Forum for Public Health, Dhaka, Bangladesh
June 9, 2014
Abstract
The hypothesis that participation in decision-making by intended beneficiaries of social pro-
grams improves the outcomes of those programs has long been influential in the academic litera-
ture and in policy. This paper presents the first experimental evidence on the effect of transferring
decision-making authority to targeted beneficiaries on the impact of a project to provide a local
public good. We randomly assigned participatory and non-participatory decision-making structures
to communities who received an otherwise identical intervention, a package of technical advices
and subsidies to provide safe drinking water sources. Participation in decision-making resulted in
larger reported increases in access to safe drinking water, but only when we imposed rules on the
decision-making process that were designed to limit the appropriation of project benefits by elite
or influential groups or individuals. Villages in which communities participated in decision-making
under rules designed to prevent appropriation reported a significantly greater increase in access to
safe drinking water (an increase of 25%) relative to villages in which project staff took decisions
(14%). In villages in which the communities participated in decision-making without imposed rules,
the change in access to safe drinking water was the same (14%) as in villages in which project staff
took decisions. We conclude that the rules we applied to limit appropriation – minimum represen-
tation requirements and decision by unanimous consensus – were effective in accomplishing their
objective.
∗E-mail: [email protected]. Funded by the National Science Foundation Grant 0624256 DHB: Decentraliza-tion and Local Public Goods: How Does Allocation of Decision-making Authority Affect Provision?. Thanks to all atNGO Forum for Public Health for their kindness and hospitality as well as their collaborative efforts, in particularExecutive Director Mr S.M.A. Rashid. Kabid Ahmed, Md. Shamsul Karim, Fatema Tuz-Zohora, and the indomitableMd. Rofi Uddin (Robi) provided excellent research assistance. Thanks to seminar participants at Columbia Universityfor helpful questions and comments.
1
1 Introduction
The hypothesis that participation in decision-making by intended beneficiaries of social pro-
grams improves the outcomes of those programs has been influential in the academic literature and
in policy for some time (e.g. Stiglitz, 2002; World Bank, 2004). Advocates of the policy argue that
involving communities in project decision-making has multiple benefits: improving project target-
ing, by drawing on information available to the community but not to outsiders; increasing ‘buy–in’
and generating a ‘sense of ownership’ of the project, thereby improving long-term management
and increasing maintenance of program assets; and promoting transparency and accountability in
project delivery. However, programs in which communities participate in decision-making may be
more susceptible to the ‘capture’ of project benefits by elite or influential community members1.
Much of the early evidence in support of this hypothesis was based on cross-sectional analyses2,
case studies3, or was simply anecdotal. Since the choice of a decision-making structure is likely to
be otherwise correlated with project, community and implementing agency characteristics, identifi-
cation of causal effects is difficult and sensitive to critical assumptions. This paper presents the first
experimental evidence on whether transferring decision-making authority to intended beneficiaries
affects the impact of a project to provide a local public good.
We randomly assigned different decision-making structures to communities who received an
otherwise identical intervention, a package of subsidies and technical advice to provide up to three
sources of safe drinking water. Many rural Bangladeshi communities currently use sources of
water that are susceptible to arsenic or, less commonly, bacterial contamination. Arsenic-safe
drinking water sources are relatively expensive and the vast majority of households cannot afford
to obtain them for themselves. As a result, the sources must generally be provided at a community
level. The random assignment ensured that the communities in which we implemented the project
under different decision-making structures were comparable in terms of all other characteristics,
allowing us to draw causal inferences about the impacts of the decision-making structures on project
1See Mansuri and Rao (2013) for a comprehensive review.2Examples include: Isham, Narayan, and Pritchett (1995), Sara and Katz (1997), A. Khwaja (2004), Fritzen
(2007), A. I. Khwaja (2009).3Examples include: Kleemeier (2000), Fung and Wright (2003), Rao and Ibanez (2005).
2
outcomes.
The three decision-making structures assigned included a non-participatory decision-making
structure and two participatory decision-making structures. In the non-participatory decision-
making structure, project staff took all decisions, based on information provided by the community.
In the first participatory decision-making structure, the community took all decisions using their
own internal decision-making processes. This process was designed to approximate the way in which
‘participation’ is implemented by organizations which place a high value on minimizing interference
with local institutions. In the second, we imposed rules on the decision-making process. Under
these rules, the community took all decisions by unanimous consensus at a meeting organized by
project staff, with requirements imposed for representation of women and the poor. This process
was designed to approximate the way in which other organizations implement ‘participation’, which
actively aim to broaden participation and reduce elite influence in decision-making.
Under all decision-making models, we retained an important participatory component. After
decisions were taken, all treatment villages were required to contribute between 10 and 20% of
the total cost of water source installation. The communities then had to decide whether or not
they would contribute, and how this contribution would be raised. We therefore identify the
effects of participation and decision-making over and above the effects generated by any financial
contribution.
Overall, the intervention led to an increase in reported access to safe drinking water of 16%
relative to a control group. The average treatment effect rises to 18%, compared to a matched
control group, when we exclude a subset of villages in which the only feasible technology for
providing arsenic-safe drinking water year-round was an arsenic iron removal plant (AIRP). This
technology has experienced issues with reliability and effectiveness in the past (Hossain et al., 2005)
and our experience suggests that communities strongly prefer tubewells to AIRPs. The treatment
effect in the villages in which AIRPs are the only technically feasible option is not statistically
different from zero, compared to a matched control group.
The increase in access to safe drinking water was higher in villages in which the community
took decisions and in which decision-making rules were imposed (22% in all villages; 25% if we
exclude the AIRP villages) compared to the villages in which project staff took decisions (13%;
14% if we exclude the AIRP villages). However, no differences were observed between the increases
3
in access to safe drinking water when the community took decisions without the imposition of
decision-making rules (14%; 15% if we exclude the AIRP villages) and when project staff took
decisions. The difference between the change in reported access to safe drinking water in villages in
which the community took decisions under imposed rules and the remainder of the treated villages
is significant when we remove the villages in which AIRPs were the only option from the analysis.
Since the treatment effect is zero in these villages regardless of the structure under which decisions
were taken, including them in the analysis is not informative with regards to a comparison between
decision-making structures.
We installed an average of 2.1 arsenic safe water sources in each of 127 treatment villages.
We installed a slightly larger number of wells in villages in which the community was involved in
decision-making (2.2 across both participatory decision-making structures) compared to those in
which project staff took decisions (2.0). However, the differences are not statistically significant.
Under the non-participatory structure, project staff were instructed to propose locations for water
sources in public spaces wherever feasible in order to facilitate access to the sources. Under the
participatory structures, communities were more likely to locate the water sources on private land.
We installed 1.9 sources per village on public land when project staff made decisions, and 1.3
when communities took decisions. A significantly smaller number of individuals contributed money
towards the water sources in the communities which took decisions without any imposed rules (5
individuals per village), when compared to the other two models (9 individuals).
Recent experimental studies have explored several aspects of ‘participation’. One influential
group of experimental studies examine the impact of a participatory or ‘community-driven’ de-
velopment project compared to a control group which does not receive any intervention (Fearon,
Humphreys, & Weinstein, 2009, 2011; Humphreys, de la Sierra, & van der Windt, 2012; Casey,
Glennerster, & Miguel, 2012)4 . The most closely related studies to this one explore variations in
how participatory decision-making processes are implemented in projects to provide a local public
good, conditional on implementing some kind of participatory decision-making approach: Olken
4Another related group of experimental studies focuses on varying requirements for participation of women ina particular decision-making process (Humphreys et al., 2012; Casey et al., 2012) or institution (Chattopadhyay &Duflo, 2004). Other related experiments examine changes in incentives to participate in school monitoring committes(Banerjee, Banerji, Duflo, Glennerster, & Khemani, 2010) or changes to the institutional structure of those committees(Pradhan et al., 2014); participation in monitoring of road construction projects (Olken, 2007) and public health careproviders (Bjorkman & Svensson, 2009); participation in project targeting for household-level interventions (Alataset al., 2013); and dispute resolution training to improve informal institutions (Blattman, Hartman, & Blair, 2014).
4
(2010) compares decisions taken at representative-based meetings to those taken by direct election-
based plebiscites; Beath, Christia, and Enikolopov (2013) compare decisions taken by secret ballot
referenda to those taken at consultation meetings. Our study differs from these two studies in three
ways. First, while these studies infer that participation in decision-making does influence decisions
taken or other outcomes, since changes in the participatory process alter these outcomes, they do
not directly measure the effect of introducing participation in decision-making itself. By including a
treatment group in which the project is implemented, under otherwise identical conditions, without
community participation in decision-making, we are able to measure the effect of introducing com-
munity participation in decision-making. Second, the two participatory decision-making processes
we compare differ from those that these studies consider; neither decision-making by consensus nor
decision-making without any imposed rules have previously been explored. Finally, the preced-
ing studies have so far only reported results on how changing the participatory process alters the
decisions taken, while we are also able to report data on the project impacts5.
Our results confirm that involving communities in decision-making can lead to greater project
impacts in terms of number of projects successfully completed and changes in reported access to
safe drinking water. However, the results also suggest that devolving decision-making authority to
the community without measures to avoid co-option of the decision-making process by influential
groups or individuals can lead to an increased incidence of elite capture. In our case, the number
of safe water sources constructed increases without any reported increase in access to safe drinking
water.
The paper is structured as follows. Section 2 describes the setting, the experimental design and
the data; section 3 describes the results, and section 4 concludes.
2 Setting, Experimental Design and Data
2.1 Arsenic Pollution Problem in Bangladesh
The context for this study is the arsenic contamination problem in rural Bangladesh. In the
1970s and early 1980s, many international agencies promoted the use of groundwater — water from
5Olken (2010) reports results on decisions taken, consistency with preferences of different groups within thecommunity, knowledge about the project, and satisfaction with the project processes; Beath et al. (2013) reportresults on project satisfaction, and consistency of decisions with ex-ante preferences.
5
wells — as a safer alternative to surface water — collected from ponds or rivers — which is often
contaminated by pathogens. At the time, noone had realized that groundwater in the region some-
times has naturally occurring high concentrations of arsenic. Arsenic contamination is not readily
detectable in water, and symptoms of arsenic poisoning only appear after years of exposure and
accumulation in the body. Information about high concentrations of arsenic in tubewells emerged
only in the mid-1990s. By that time, the damage was done; the resulting epidemic of diseases asso-
ciated with arsenic exposure has been described as ‘the largest poisoning of a population in history’
(Smith, Lingas, & Rahman, 2000). In 2008, when this project began, UNICEF estimated that 20
million people were still using water from wells with arsenic concentrations above the Bangladeshi
standard, which is itself five times higher than the WHO standard (UNICEF, 2008).
Creating access to safe drinking water in the presence of arsenic contamination presents a prob-
lem of providing a local public good. The great majority of drinking water sources in Bangladesh
are privately owned, including almost all tubewells that have high concentrations of arsenic. Tech-
nologies to provide water with low concentrations of arsenic are considerably more expensive, and
entail high fixed costs. Only the richest households can afford to purchase these sources them-
selves. For most households, they must be provided at the community level, at which high fixed
costs can be shared among many people. As a result, communities who wish to improve access to
safe drinking water must typically solve a collective action problem.
Several technologies are available to provide arsenic-safe drinking water, of which deep tubewells
are the most common in rural Bangladesh. Deep tubewells draw water from deep aquifers (approx-
imately 700-800 feet below ground level) that have low concentrations of arsenic. Standard deep
tubewells are relatively expensive to install, but easy to use and maintain, and replacement parts
are readily available. In some areas, arsenic safe water is available at lesser depths of approximately
300-400 feet. In these areas, shallow tubewells can provide arsenic-safe drinking water, at a lower
installation than deep tubewells. Shallow tubewells are otherwise very similar to deep tubewells in
terms of functionality, maintenance requirements and ease of repair.6 In some areas, there is con-
siderable seasonal variation in water pressure in the aquifer and standard deep tubewells may not
provide year-round access to safe drinking water. An alternative design — the deep-set tubewell
6During the study implementation period, information emerged about a problem of manganese contamination inshallow tubewells. As a result, we replaced shallow tubewells we had already installed free of charge with alternativetechnologies, if they tested positive for manganese.
6
— can provide year-round access to safe drinking water in these areas. The pumping mechanism
in the deep-set tubewell is installed below the surface of the ground, as opposed to on the surface
in the standard design. This means that the deep-set tubewell is more expensive and more difficult
to repair in case of failure than the standard deep tubewell, but it is equally convenient and easy
to use.
In some areas, there is no accessible arsenic-safe aquifer – for example, where an intermediate
layer of rock cannot be penetrated using local drilling techniques – and therefore it is not feasible
to install tubewells. An alternative technology is the arsenic iron removal plant (AIRP). AIRPs
remove arsenic by oxidation and filtration. They are more expensive, larger and significantly more
difficult to operate and maintain than tubewells, and our experience suggested that communities
strongly preferred tubewells.7 As a result, we will throughout the paper report treatment effects
by the type of feasible technology – AIRPs or tubewells – as well as the overall treatment effect.
2.2 Experimental Design
The project intervention consisted of a package of technical advice and subsidies for the provision
of up to three safe drinking water sources per community. We carried out the interventions between
2008 and 2011, in partnership with a Bangladeshi non-governmental organization (NGO), NGO
Forum for Public Health. NGO Forum for Public Health is a well-established actor in the water
and sanitation sector with more than 30 years experience in the field.
We conducted our study in communities located in two upazilas (subdistricts): Gopalganj,
about 60 miles southwest of Dhaka, and Matlab, about 30 miles southeast of Dhaka. We focused
on these sites because of the severity of the arsenic contamination problem in the area more than
80% of pre-existing tubewells were arsenic contaminated and because the sites had not yet received
other interventions to address the problem. We studied 250 villages, equally split between the two
upazilas, and ranging in size from a minimum of 7 households to a maximum of 1103, with the
median size 170 households.8 Before interventions began, we carried out an information campaign
about the arsenic problem, to ensure that all villages were initially equally well informed about the
7Where tubewells were not feasible, we also offered communities the opportunity to install rainwater harvestingsystems or a pond sand filter, but since no community selected either of these options, we do not describe themfurther in the paper. Both technologies have limitations with respect to tubewells or AIRPs
8Data on arsenic contamination of pre-existing tubewells and village size was drawn from the Bangladesh ArsenicMitigation Water Supply Project.
7
arsenic problem.
Of the 250 villages studied, we assigned 100 to a control group who did not receive the inter-
vention. 126 villages received the intervention. We initially assigned a further 24 villages to receive
the intervention who eventually did not receive the intervention, due to changes in the costs of
providing safe water sources over the course of the project. We originally assigned one other village
to treatment, but project staff determined before the project began that there were no feasible
available technologies to provide safe drinking water in the community, because no arsenic safe
aquifer was accessible, and arsenic concentrations in the shallow groundwater were too high for
removal with an AIRP. There was one other village in which we determined after we began the
intervention that there were no feasible available technologies to provide safe drinking water.
The original protocol for selection of treated villages was random, which should have resulted
in treatment and control groups which were comparable at baseline. However, we later established
that the project director at the time, who was later removed from the project for unrelated reasons,
did not follow the original protocol when he implemented the division of the villages into control
and study villages, and he included all villages in the southern area of Matlab in the treatment
group. Villages in South Matlab have much lower access to safe drinking water than the average
village in the sample, meaning that overall the treated group had significantly lower access to safe
drinking water at baseline than the control group.
Table 1 confirms that this resulted in statistically significant differences between control and
treatment groups. Treated villages had reported lower access to safe drinking water, and were less
likely to have changed their source of drinking water because of the arsenic contamination problem
in the last five years. In Table 1, we show baseline summary statistics and randomization checks
for villages by treatment status. The table shows the mean and standard errors for a selection of
baseline variables which measure baseline access to safe drinking water, factors that might influence
the ease of providing safe drinking water, and community-level variables that might influence the
likelihood of a successful collective action. In column 2), we test whether the difference in means
between treated and control villages is statistically significant. The p-values are derived from
Ordinary Least Squares (OLS) regressions with the following structure:
8
Yi,v = α+ βItreated,v + εi,v (1)
where Yi,v is the value of a variable in household i in village v and Itreated,v is an indicator which
is one if village v was treated and zero if village v was not treated. If the treatment was randomly
assigned, the coefficient β should be zero as assignment to treatment should not be correlated with
any baseline characteristics of the village. The p-values test whether the coefficient is equal to zero.
Since treatment was assigned at the village level, but we collected data at the household level, it
is important to account for within-village correlation in variables. Within-village correlation implies
that it is more likely that differences between mean outcomes in treated and control villages arise
due to chance, than if we had been able to assign treatment at the household level. In order to
ensure that the statistical analyses we carry out make the correct inference about whether or not
a result is likely to be due to chance or not, we follow Angrist and Pischke (2009) and cluster
standard errors at the village level.
Columns 3) to 5) of Table 1 show that we can correct for the bias induced by the failure of
randomization by three methods. First, we can drop South Matlab from the sample. Second, we
can create a synthetic treatment variable generated at random in South Matlab, and equal to the
treatment variable elsewhere 9. This synthetic variable re-assigns a fraction of the villages in each
treatment group in South Matlab to control. Third, we can use this synthetic treatment variable to
instrument for treatment. In columns 3) to 5), we report the difference in means between treated
and control villages under these three approaches, after accounting for the different proportions
of treated villages in Gopalganj and Matlab, because differences in treatment and control groups
otherwise reflect differences between these areas. We estimate the difference in means using an
equation similar to Equation 1, including indicators for Gopalganj and South Matlab. In each case
we show that no significant differences remain between treatment and control populations.
Since the non-random selection of treatment villages in South Matlab may have introduced bias
into our estimates of treatment effects, we therefore report both OLS results and results where
treatment is instrumented using the synthetic treatment variable (the IV results).
9Ideally, we would have used the original random assignment to treatment rather than this synthetic alternativebut we have not been able to recover the initial, randomly assigned treatment lists.
9
Decision-making structures
Project staff implemented the intervention under one of three decision-making structures. The
necessary decisions included if, how and where to install; and how to manage, each safe drinking
water source. In all cases, project staff ensured that all decisions made were technically appropriate.
Table 2 summarizes the main features of the different decision-making structures. We describe the
three models in more detail in the following paragraphs.
The decision-making structures included one non-participatory structure, the Top-Down model
(TD). Under this model, project staff took all project decisions, after an extended (typically 2-
day) period of information gathering. The information gathering process consisted of participatory
mapping of the village with members of the community, focusing on the locations of households
and safe and unsafe sources of drinking water, cross-checking information with various community
members. Project staff then proposed sites for safe drinking water sources, prioritizing locations
with the highest density of households not already served by safe drinking water sources, choosing
public locations wherever possible, and convenient locations where no suitable public land was
available. Staff then organized and publicized a community meeting at which they presented the
proposed locations. This model was designed to approximate the ‘traditional’ approach to decision-
making about local public goods in which decisions are taken by a centralized organization, such
as local government or an NGO.
The decision-making structures also included two participatory structures, in which decision-
making authority was devolved to the community. Under the ‘pure’ Community Participation (CP)
model, project staff visited the community to arrange a meeting at a site and time of the commu-
nity’s choosing. At the meeting, project staff explained the project rules and announced that they
would return to the village after a few days to find out whether they wanted to participate in the
project, and if so, which sites they had chosen. Sites that were not technically appropriate were
rejected, but otherwise the community’s decisions were final, conditional on raising the commu-
nity contribution. We did not directly observe the decision-making process used, but communities
reported to us that they took these decisions in a variety of ways including open meetings (some-
times but not always including women), meetings at a mosque, or closed-door meetings of village
elites. This model was designed to approximate the way in which some organizations implement
10
community participation in practice, avoiding interference with a community’s internal hierarchies
and decision-making processes .
Under the second participatory decision-making structure, the NGO-Facilitated Community
Participation model (NGO), we imposed rules about how decisions should be taken. Project staff
initially organized a series of separate small group meetings with men and women who the com-
munity identified as poor and non-poor. At these small group meetings, project staff explained the
project rules and emphasised the right all individuals would have to participate in the decision-
making process and benefit from the interventions. These meetings were followed by a community
meeting, at which both men and women, and poor and non-poor, had to be represented. The
community proposed and selected water source locations by unanimous consensus at the meeting,
in the presence of project staff and with their active facilitation. If the community could not reach
a consensus at the first meeting, a second and in some cases subsequent meetings were organized.
This model was designed to approximate the way in which other organizations implement commu-
nity participation, with project staff playing a strong facilitatory role, and rules imposed that are
intended to reduce the likelihood that influential groups or individuals co-opt the decision-making
process.
Before installing a safe drinking water source, we required the community to contribute between
10% and 20% of its cost, depending on the technology installed. Table 3 shows the cost of installing
each of these technologies and the community contribution that we required. The difference in
required community contributions reflects the difference in cost of the selected technology. We also
scaled the community contribution so that the subsidy could be either concentrated on one water
source or spread between up to three water sources. The price per water source therefore increased
as more water sources were installed in the village. Budget constraints meant that when the best
feasible technology was one of the more expensive alternatives, we were only able to offer up to two
water sources.
After the initial decision-making process, project staff gave the communities up to twelve weeks
to raise the funds for the community contribution. Construction of the safe drinking water sources
began as soon as the community had raised their contribution. If after twelve weeks the community
had not raised their contribution, construction of the safe drinking water sources did not go ahead.
We initially intended the decision-making structures to apply to decisions about who contributed
11
to the community contribution, but this proved impossible to enforce. However, project staff did
propose a list of contributors at the Top Down model meetings, and communities did agree a list
of contributors at the NGO-Facilitated Community Participation meeting.
We randomly assigned the decision-making structures to the communities who received the
intervention. Of the 126 treated villages, we initially assigned 42 to each decision-making model.
We replaced the village in which we determined before beginning the project that there was no
feasible safe drinking water technology with another village, randomly drawn from the villages
which we had initially assigned to treatment but in which we had not carried out the intervention
due to budget constraints. As a result 43 villages were assigned to the Top-Down model.
Table 4 shows that the villages assigned to each decision-making model were comparable at
baseline to the villages assigned to the other decision-making models. We test whether the difference
in variable means between villages in which the project was implemented under a given decision-
making structure and the remainder of the treated villages is statistically different from zero. The
p-values in the table are therefore derived from OLS regressions similar in structure to Equation
1 but the indicator Im,v is one if village v received treatment under decision-making structure m,
and zero otherwise:
Yi,v = α+ ΣβmIm,v + εi,v (2)
Only the treated villages are included in the regressions in Table 4. We do not use the control
group for comparison in this case because the results in Table 1 already confirm that the treated
villages are not directly comparable to the control villages.
We compare 15 variables across the 3 decision-making structures, resulting in a total of 45
tests. In 43 of these tests we fail to reject at the 10% level the null hypothesis that there is
no difference in means between groups treated under one decision-making structure and the other
treated villages. In 2 tests we find statistically significant differences between the mean of a variable
in villages treated under one decision-making structure and in the remaining treated villages. One
test rejects this hypothesis at the 5% level, and one at the 10% level. This is consistent with what
we would expect due to chance. From these checks we conclude that there is no evidence to suspect
12
that assignment to model, conditional on treatment, was not random, as required by the project
protocol.
The same project staff – one team in Gopalganj and one team in Matlab – implemented the
project under all three decision-making structures. We implemented the intervention in cycles
during which project staff would complete the entire process from meeting organization to water
source installation for a group of villages, where the villages were grouped geographically for ease
of logistics. The project was initially implemented in 114 villages in 6 cycles across both upazilas.
We later added an additional 12 villages in Gopalganj when funds became available, in a 7th cycle.
Government policy had changed by the time we carried out the 7th cycle, and community
expectations that the government would provide free tubewells may have increased. We installed
fewer safe water sources under the 7th cycle, but the number installed is not significantly less than
under the first 6 cycles in Gopalganj, once we account for the feasible technology.
2.3 Data Description
We carried out a baseline survey in 2007 in 40 households in each of the 250 villages, sampled
randomly from census lists . We surveyed a total of 9797 households, as in some very small villages
there were fewer than 40 households. The baseline questionnaire included standard components of
a household survey with a special focus on social networks and social capital, and full details on
water use behavior. We also collected village-level information from focus groups.
We encountered significant problems with the data entry process after the baseline survey. First,
some of the individuals employed to enter the data in spreadsheets copied and pasted entire villages
of data, changing names and other identifiers to conceal what they had done. Data checking revealed
this problem by chance several months after data collection and entry had been carried out. When
we discovered this problem, we checked extensively for additional incidences and had the missing
data re-entered. Second, by the time we discovered this problem, termites had unfortunately
attacked the stored questionnaires, and destroyed a small percentage of the questionnaires. As
result, we are missing baseline data from 140 households from control and treated villages, since
enumerators did not initially enter the data correctly and termites then destroyed the hard copy
of the questionnaires. We do not however have any reason to think that there was any systematic
pattern to either the false data entry or the losses to termites, so the remaining baseline data should
13
still represent a randomly selected sample of the baseline population.
We carried out follow-up surveys in control and treated villages in 2010 and 2011 after we
carried out the safe water intervention, interviewing the same households that we interviewed for
the baseline survey. We did not carry out follow-up surveys in the 24 villages which were initially
assigned to treatment but in which we did not carry out the intervention. We therefore attempted
to resurvey 8,890 households from the original panel, of which we successfully re-surveyed 8630
households, representing an average attrition rate of 2.9%. The attrition rates broken down by
treatment group are as follows: 2.7% in control villages; 3.1% in treated villages. Among the
treated villages, attrition rates were 2.6% in NGO-Facilitated Community Participation villages;
3.2% in Community Participation villages; and 3.4% in Top-Down villages. The attrition rates in
treated groups and sub-groups were not statistically different from the control group, or from each
other.
We also carried out follow-up surveys in 1424 additional households in treated villages, to bring
the minimum survey coverage up to 15% of households in all treated villages (based on census data).
The additional households were randomly selected from the remaining households on the census lists
who had not been surveyed at baseline. Extending the survey coverage in this way was intended to
ensure that the survey captured the effects of the intervention in larger villages, where the three safe
drinking water sources constructed were unlikely to serve the entire community. However, the data
from these additional households is inconsistent with the data collected from the panel households,
and we have established that there were violations of the sampling protocol; in particular, in some
villages, some of these additional households were sampled from the neighbourhood of the installed
water sources, rather than from the census lists. As a result, we do not use this data in this analysis.
We also collated data on the numbers and types of safe drinking water sources installed, and
project staff kept detailed records of the implementation process, including the number of contrib-
utors in each community and the time taken to raise the community contribution. We also carried
out focus group discussions in treatment villages to obtain qualitative information about why the
project was successful in some communities and not in others.
14
3 Results
We first show how attendance at decision-making meetings varied by decision-making model to
demonstrate that the NGO-Facilitated Community Participation was marginally more successful
in including more people in the decision-making process, and in including a more diverse range of
people. We then report the project outcomes in the study villages. Projects implemented under the
participatory decision-making models were more successful in terms of installing safe water sources,
but the differences between models are not statistically significant. We were far more successful in
installing safe water sources in villages where tubewells were feasible, than in villages where only
AIRPs were feasible.
Projects implemented under the Top Down decision-making model were however much more
successful in installing safe water sources in public places, although whether an installation site is
public or not is not necessarily a good predictor of how well-used the water source will be. Finally,
we show that in villages implemented under the Community Participation model, fewer households
contributed towards the cost of installation.
We then report the average treatment effect in terms of changes in reported access to safe
drinking water for all villages. We show that overall, the project led to a 16% increase in reported
access to safe drinking water. The treatment effect was higher in villages where tubewells were
feasible (18%), and zero in villages where only AIRPs were feasible. We then show that the
treatment effect was substantially higher in villages where the project was implemented under the
NGO-Facilitated Community Participation model, and that the differences are significant when we
exclude the villages in which only AIRPs were feasible; given that the average treatment effect
is zero in these villages, including these villages is not informative with respect to a comparison
between decision-making models. Finally, we present some robustness checks on these main results.
3.1 Participation
Table 5 shows the recorded numbers and characteristics of individuals attending the main
community meetings. Overall, the mean number of participants was 30.7, with 27% of meeting
attendees female, 42% of low socioeconomic status (as recorded by project staff) and 60% with less
than secondary education.
15
More people, from a more diverse range of groups, attended meetings held under the NGO-
Facilitated Community Participation decision-making model than meetings held under the other
two decision-making structures. The number of participants was somewhat higher in NGO-
Facilitated Community Participation model meetings (33.6), and lowest in the Top Down model
meetings (28.5), but the differences between models are not statistically significant. The NGO-
Facilitated Community Participation model meetings were the most diverse in representing dif-
ferent groups, with the highest percentage of female participants and the highest percentage of
participants with less than secondary education. The Top Down meetings were the least diverse,
with the lowest percentage of female participants, and the lowest participation of participants with
low socio-economic status. However, again, not all the differences between models are statistically
significant. Nonetheless, to the extent that attendance at meetings really reflects de facto participa-
tion in decision-making, the NGO-Facilitated Community Participation model was more successful
in involving a larger number and wider range of community members in decision-making.
3.2 Project Outcomes
Table 6 shows how the decision-making model influenced project outcomes in the treated vil-
lages. On average, we installed 2.14 safe water sources in the treated villages. If we had installed
all technically feasible water sources given our project rules, we would have installed an average of
2.75 safe water sources, based on installing three sources per village in most cases, and two sources
per village where only a more expensive technology was feasible.
We offered communities the choice between all technically appropriate technologies to provide
safe drinking water, given local hydrogeological conditions. In Gopalganj, we carried out the
intervention in 70 villages. In 16 villages, AIRPs were the only feasible technology. In two villages,
no treatment was feasible, as there was a layer of impenetrable rock, and shallow groundwater was
too strongly contaminated with arsenic and iron for removal with an AIRP. In Matlab, tubewells
were feasible in all villages.
A clear preference gradient between the available technologies emerged. People chose shallow
tubewells wherever possible, followed by standard deep tubewells and deepset tubewells. The three
types of tubewell are comparable in ease of use and maintenance, but increase in cost with depth
and design complexity. AIRPs were the least preferred option by a wide margin. There were 16
16
villages, all in Gopalganj, where the only type of water source that could be installed was an AIRP,
meaning that we could have installed a total of 32 AIRPs. We were only successful in installing 5
AIRPs during the course of the project, a success rate of approximately 16%. In comparison, in
the remaining villages in Gopalganj — in which tubewells were feasible — we installed 79% of the
maximum number of wells we could have installed under our project rules. The reasons given by
the communities for rejection of the AIRPs were that they took up too much space, required too
much work to operate and maintain, and were not perceived to be reliable or trustworthy. When
we consider only the villages in which tubewells were feasible, the average number of water sources
constructed rises to 2.45 out of a maximum possible 2.85.
The rejection of AIRPs did not seem to be a direct function of the price of the technology.
However, in Matlab, in the 10 villages where only deep-set tubewells could be installed (which are
comparable in price to AIRPs, and for which we required the same level of community contribution),
we installed on average 90% of the maximum feasible number of water sources, compared to an
average of 89% in all other villages in Matlab (where either deep tubewells or shallow tubewells were
feasible). However, we cannot comment on what would have happened if we had offered AIRPs at
a lower price.
We installed 10% more water sources in the villages in which communities participated in
decision-making than in the villages in which project staff took decisions, as shown in column 1).
Installing more water sources is one measure of success of the project, but it may not translate
into increased access to safe drinking water if the sources are not fully accessible to the community.
However, the differences are not statistically significant. In Table 6, we assess whether differences
in project outcomes across models are statistically significant using OLS regression for the following
equation:
Yv = βNGOINGO,v + βCP ICP,v + βTDITD,v + εv (3)
We then test pairwise equality of the coefficients βNGO, βCP and βTD. The differences between
the number of water sources installed under the different decision-making models are attenuated
further in both magnitude and significance when we consider only the villages in which tubewells
17
were feasible.
We installed more water sources in public spaces, as recorded by our project staff, under the non-
participatory Top-Down model. Public spaces were defined to include communal land, open spaces,
areas beside roads, and institutions such as mosques or schools, as opposed to privately owned land.
Under the Top-Down model, project staff had a specific mandate to install water sources in public
places. The differences are strongly significant with respect to both the participatory decision-
making models. Water sources installed in public places may be accessible to a larger number
of people. However, space that is appropriate for water source construction is quite strongly
constrained in villages in this region, and the most convenient location for a water source may
not necessarily be located on public land. This is primarily because land that is not vulnerable to
flooding is relatively scarce, and safe water sources cannot be installed on land that is vulnerable
to flooding because of potential contamination.
Fewer people contributed to raising the community contribution in the unregulated Community
Participatory model than under the other two models. The difference is significant with respect to
both the other two decision-making models. A small number of contributors may be efficient, as
some community members will have a greater ability to contribute than others. However, it may
also be indicative of a high degree of influence over the decision-making procedure, which may not
be efficient if used to co-opt project benefits for private use.
Overall, the number of contributors was relatively low in all cases, considering that the median
village size was 170 households. In villages where we successfully installed at least one safe water
source, the mean number of contributors per water source installed was 5.1 in NGO villages, 2.3 in
CP villages and 4.0 in TD villages. There was only one contributor per safe water source installed
in 34% of the NGO villages, 56% of the CP villages and 45 % of the TD villages.
3.3 Reported Project Impact
We primarily measure access to safe drinking water based on an outcome variable which mea-
sures whether or not the household reports using safe drinking water. The indicator is based on
the source of water that the household identifies as being its most important source of water for
drinking and cooking. The indicator for reporting use of safe drinking water is constructed as being
equal to one where the household reports using a source of drinking water that is safe from both
18
bacterial and arsenic contamination, and zero when they report that the source is unsafe, if they
don’t know whether it is safe or not, or if it is a source that is vulnerable to bacterial contamina-
tion e.g. a dug well or surface water. Further details regarding the construction of this variable is
included in Appendix A.
We report the average overall treatment, relative to the full control group, but we also break
down the treatment effect by whether tubewells or only AIRPs were feasible. There is strong spatial
correlation between locations where only AIRPs are feasible, reflecting the extent of the rock layer
overlaying the deep aquifer. Since other village level characteristics are also spatially correlated,
there are as a result some differences on baseline characteristics between villages in which tubewells
were feasible and villages in which only AIRPs were feasible in Gopalganj.
When we report effects for villages in which a specific technology was feasible, we use a matched
control group, because we do not observe which technologies are feasible in the control villages.
Using a matched control group removes these differences (see Appendix Table B1). However, in
robustness checks, we show that the results for tubewell villages are not sensitive to using a different
matched control group, or to simply using the full set of control villages. The construction of the
matched control group exploits spatial correlation in the location of villages in which AIRPS were
the only feasible technology. Details of the construction of the matched control group are given in
Appendix A.
Average treatment effect
In Table 7 we show reported access to safe drinking water at baseline (Panel A) and follow-up
(Panel B), and the resultant change in access (Panel C). We show results for all villages in columns
1) to 3); in all villages in which tubewells were feasible in columns 4) to 6); and in those villages
where only AIRPs were feasible in column 7).
Columns 1) and 4) show the OLS results, which as previously discussed show that treated vil-
lages have worse access to safe drinking water at baseline. Columns 2) and 5) show the results
dropping South Matlab, where the randomized assignment to treatment was not correctly imple-
mented. Columns 3) and 6) show IV results, using a synthetic assignment to treatment variable
in South Matlab. EIther of these approaches removes the baseline differences in reported access to
safe drinking water. In Gopalganj, there were no problems with random assignment to treatment,
19
so we only report one set of results.
To estimate the results in Panels A) and B), we use data from all panel households and estimate
the following equation:
i,t= α+ βItreated,v + εi,t (4)
where i is a household in village v and Yi,t is the access to safe drinking water at baseline or
followup.
Panel A) shows baseline access to safe drinking water. This panel repeat the comparisons shown
in Table 1, and demonstrates the validity of the approaches used for compensating for the failure of
random assignment. Panel B) shows the follow-up comparisons. As a result of the initial differences
between treatment and control villages, the differences between follow-up and control villages are
not statistically significant at follow-up in the OLS analysis, but are clearly evident when we correct
for the failure of random assignment.
To estimate the change in access to safe drinking water between baseline and followup (Panel
C), we estimate a first difference equation for all households for which we have both baseline and
follow-up data, as follows:
∆Yi = Yif − Yib = α+ βItreated,v + εi (5)
where ∆Yi is the change in access to safe drinking water between baseline and followup. With
two time periods, the first difference analysis is directly equivalent to including household fixed
effects. As before, we cluster standard errors at the village level to account for within-village
correlation in outcomes.
Panel C) shows the change in reported access to safe drinking water. The OLS and IV results
are almost identical, suggesting that although assignment to treatment was not random in all areas,
it was not correlated with trends in reported access to safe drinking water. The estimated average
treatment effect is 16% overall, and 18% in villages in which tubewells were feasible. There is no
20
treatment effect in the AIRP villages.
For these results and the remainder of the results in this section, we use survey weights which
ensure that each village counts equally in the analysis. Where part of the data for a village was
lost through the baseline data entry problems, the baseline weights compensate for these losses, as
there is no reason to think that the lost data introduces any bias to the estimates of a variable in
the village. We do not introduce compensatory weights for migration, but attrition rates were low
overall, so this is unlikely to influence the results.
Treatment effect by decision-making model
Table 8 repeats the analyses shown in Table 7, with the estimated effects broken down by the
decision-making model under which we implemented the project. Once again, Panel A) shows
baseline access to safe drinking water, Panel B) shows access at follow-up, and Panel C) shows the
change in access. Columns 1) to 3) show the results in all villages; columns 4) to 6) show the results
in villages where tubewells were feasible; and column 7) shows the results where only AIRPs were
feasible. In columns 2) and 5), we correct for failure of random assignment to treatment by dropping
South Matlab; and in columns 3) and 6) we correct by instrumenting for the model assigned by the
interaction between an indicator for the implemented model, and the synthetic treatment variable.
In Panel A) and B), we use equations with the following structure:
Yi,t = α+ βNGOINGO,v + βCP ICP,v + βTDITD,v + εi,t (6)
where Yi,t is reported access to safe drinking water at baseline or follow-up in household i in village
v, and I is an indicator whether the village was treated under a given decision-making structure.
Panel A) shows that villages treated under each decision-making model differ from the control
villages (as a result of the failure of random assignment to treatment), but that the villages treated
under each model are comparable to each other. Columns 2), 3), 5) and 6) also confirm that
the strategies for correcting for the failure for random assignment to treatment are also effective
for removing significant baseline differences between villages treated under a given model and
the control villages as a whole. This is consistent with the results shown in Table 4, in which
21
we showed that conditional on treatment, villages assigned to different decision-making models
were comparable on baseline statistics. However, note that the magnitude of differences between
treatment groups is quite large with respect to the treatment effects we estimate. In particular,
36% of households in NGO Facilitated Community Participation villages report having access to
safe drinking water at baseline in comparison to 41% in Community Participation villages and 44%
in Top Down villages.
Panel B) shows that, in the specifications when we correct for the failure of random assignment to
treatment, there are significant treatment effects under all decision-making models. The estimated
effects using only follow-up data are largest for the NGO Facilitated Community Participation, but
the differences across models are generally small. However, note that the pattern of access to safe
drinking water is reversed: villages in which the project was implemented under NGO Facilitated
Community Participation model have the lowest access to safe drinking water at baseline, and the
highest access to safe drinking water at follow-up.
In Panel C), we show the results for the change in reported access to safe drinking water. We
find that the reported increase in safe drinking water is greatest in NGO model villages, and the
reported increase in safe drinking water is almost exactly equivalent in TD and CP model villages.
We estimate a first difference regression of the change in reported access to safe drinking water
using the following equation:
∆Yi = Yi,f − Yi,b = α+ βNGOINGO,v + βCP ICP,v + βTDITD,v + εi (7)
The estimated coefficients are extremely consistent, with the estimated increase in access to
safe drinking water between 21% and 22% in NGO-Facilitated Community Participation model
villages (24% and 25% in tubewell villages); between 13% and 14% in Community Participation
model villages (between 13% and 15% in tubewell villages) and between 11% and 13% in Top
Down model villages (between 12% and 15% in tubewell villages). None of the models shows any
significant treatment effect in AIRP villages.
The size of the effect is economically quite important, as the treatment effect almost doubles.
However, the differences between models are just below the threshold of statistical significance
22
when all villages are considered. The difference in size of the treatment effect between the NGO-
Facilitated Community Participation model villages and the other treated villages is statistically
significant when we exclude the villages in which only AIRPs were feasible. Since the treatment
effect was zero in these villages, including these villages reduces all the estimated model-specific
treatment effects, making it more difficult to distinguish between them, and introduces noise that
is not informative with respect to a comparison between the decision-making models.
We do not include data from the additional households that we surveyed at follow-up, because
of inconsistencies between the additional households and the panel households, and because of
concerns that the sampling protocol may not have been implemented consistently. Including data
from the additional households surveyed at followup increases the magnitude and the statistical
significance of the difference between the NGO model villages and the other treated villages.
3.3.1 Robustness Checks
In Table 9, we show the effects of changing the main specification on the estimates of the size of
the treatment effects under the different decision-making models. Unless specified, we focus only on
the tubewell villages, where the differences between models are statistically significant. For brevity,
we only report the p-values of the tests of interest, given the main results: whether the results from
the NGO-Facilitated Community Participation model villages are equivalent to the results from
the other villages.
In Column 1), we show that the results are similar when we consider only the treated villages,
and do not include the control villages in the analysis. The results are given as OLS as within the
treated villages, assignment to model was random.
In columns 2) and 3) we show the results by upazila. The treatment effects are larger in
Gopalganj, and the differences between models are more pronounced. However, the pattern of
results is also consistent for Matlab in that the increase in access to safe drinking water is largest
in the villages treated under the NGO-Facilitated Community Participation model.
In columns 4) and 5), we show that the results are not sensitive to how the matched control
group is constructed. Column 4) shows that an alternative matched control group, generated
by assigning villages in Gopalganj to AIRP and tubewell-matched groups at random, given the
probability of AIRP/tubewell feasibility in their neighbourhood, yields almost identical results.
23
Column 5) shows that using the full control group also yields similar estimates.
In columns 6) to 8), we include additional controls. We include the interactions between the
control variables and the treatment variables, since otherwise including control variables results
in bias in the estimated treatment effect (Freedman, 2008)10. In column 6), we use the full set of
treated villages and control for the best available technology. We do not include the control villages
as we do not directly observe the best available technology in these villages. The coefficients are
very similar.
In column 7), we include a quadratic function of village size, and its interactions with the
treatment indicators. Allowing for heterogeneity in the treatment effect by village size increases the
size of the estimated effect in NGO-Facilitated Community Participation. This reflects substantial
heterogeneity in the overall effect, and the effect by model, across village size. Figure 1 shows the
heterogeneity in the overall treatment effect. The effect size decreases with village size, reflecting the
fact that the intervention was limited to install a maximum of three safe water sources, regardless
of village size, and possibly, the increasing difficulty of solving a collective action problem with
a group of increasing size, as theory predicts (Olson, 1971). No treatment effect is detectable in
villages with more than 500 households, although the number of villages in this group is small (8
treated villages and 10 control villages). Figure 2 shows the heterogeneity in the treatment size by
decision-making model:, the effect sizes decreases for each decision-making model, with the effect
size greatest under the NGO Facilitated Community-Participation model over the entire range of
sizes at which treatment effects are observed. We use the full set of treated and control villages in
this column, and column 8.
In column 8), we include a quadratic function in total household assets and its interactions with
the treatment indicator. The results are similar to the main specification.
In columns 9) and 10) we report results using alternative measures of access to safe drinking
water, described in detail in Appendix 4. In column 9, we report results using a measure of
whether the source the household was using could be verified to be safe. To verify safety of a
tubewell, enumerators inspected the tubewell that the household reported using if it was less than
10We include interactions with the raw treatment indicator, rather than the synthetic treatment indicator usedfor the first stage, as we are not clear which is the correct procedure in the context of an instrumental variablesregression. However, the change in the results is minor if we instead control for interactions with the synthetictreatment variables.
24
5 minutes walk away, and recorded whether it was marked red (unsafe), green (safe) or unmarked.
The increased treatment effects are larger across all models, partly because our intervention also
increased the fraction of tubewells that were verifiably safe, and also because we did not collect
this data from all villages at baseline. In column 10), we report results using a measure than uses
the verified data when it is available, and the reported data when enumerators were not able to
verify the safety of a source. The results are broadly consistent with the results which use only the
reported measures.
4 Conclusions
This study has provided the first experimental evidence to support the claim that delegating
decision-making authorities to communities in projects to provide local public goods can improve
outcomes and increase reported impact. In villages where we implemented a project to provide safe
sources of drinking water under a participatory decision-making structure (the NGO Facilitated
Community Participation model), we installed a slightly larger number of safe drinking water
sources (0.2 more sources) but obtained a 9% higher increase in access to safe drinking water,
than under a non-participatory decision-making structure (the Top-Down model). Under this
model, community members took key decisions by unanimous consensus at a community meeting
with minimum representation requirements. These results are broadly consistent with evidence
accumulated in the past through practitioner’s experience and cross-sectional analysis, but this is
the first time that experimental evidence has been available to test the hypothesis that participation
in decision-making has a positive impact on the result of social programs.
However, the study also suggests that these benefits may not be realised if protective measures
are not put in place to prevent the decision-making process from being co-opted by influential
groups or individuals. Under the ‘pure’ Community Participation model, under which communities
took decisions without imposed rules, we installed the same number of sources as under the NGO-
Facilitated Community Participation model, but we obtained a 9% smaller increase in access to
safe drinking water.
Since we did not test alternative strategies for preventing the decision-making process from
co-option, we cannot comment as to whether the method used here (imposing the requirement
25
that decisions be taken by unanimous consensus at a community meeting where all groups were
represented and conducting small group meetings beforehand to raise awareness about the project
objectives and the rights of all individuals to participate) was the most effective possible in the con-
text. We also did not delegate technical decision-making authority to the community (our project
staff determined the feasibility of any given technology and location) and therefore cannot deter-
mine whether the results would be the same or different if decision-making authority is delegated
to the community over other types of decisions.
A potential weakness of our results is that we rely on reported data, and it is possible that
participation in project decision-making may influence the way in which intended beneficiaries
report project outcomes. We have also collected data on actual use of the installed water sources by
monitoring their use directly using enumerator observations. This data is currently being analysed.
The role of the community contribution appears key in determining outcomes. The number of
contributors is low over all. Those that can contribute towards the cost of the water source may
have significant influence over the decision-making process. The number of contributors is lowest
in the pure Community Participation villages, where we find suggestive evidence of a higher degree
of elite capture. Anecdotally, project staff reported to us that in some Top-Down model villages
where community groups failed to raise the community contribution, individuals volunteered to
pay the community contribution, but only if the water source was installed on their private land.
The result of delegating decision-making authority to the community may vary substantially de-
pending on the local context, for example depending on existing inequalities within the community
or on the size and homogeneity of the group to which authority is delegated. We cannot determine
whether the results of this study would be applicable in other contexts. The study would benefit
from replication in different social and cultural contexts.
Bearing these caveats in mind, the results provide important experimental evidence regarding
an influential policy recommendation, and suggest that careful consideration should be given to the
structure of a participatory decision-making process, if the potential benefits are to be realized.
26
References
Alatas, V., Banerjee, A., Hanna, R., Olken, B. A., Purnamasari, R., & Wai-Poi, M. (2013, January).Does elite capture matter? Local elites and targeted welfare programs in Indonesia.
Angrist, J. D. & Pischke, J. (2009). Mostly harmless econometrics. Princeton University Press.Banerjee, A., Banerji, R., Duflo, E., Glennerster, R., & Khemani, S. (2010). Pitfalls of participatory
programs: Evidence from a randomized evaluation in education in India. American EconomicJournal: Economic Policy.
Beath, A., Christia, F., & Enikolopov, R. (2013, January). Direct democracy and resource allocation:Experimental evidence from Afghanistan. Centre for Economic and Financial Research at NewEconomic School Working Paper No 192.
Bjorkman, M. & Svensson, J. (2009). Power to the people: Evidence from a randomized fieldexperiment on community-based monitoring in Uganda. Quarterly Journal of Economics,124 (2).
Blattman, C., Hartman, A. C., & Blair, R. A. (2014). How to promote order and property rightsunder weak rule of law? an experiment in changing dispute resolution behavior through com-munity education. American Political Science Review, 108 (01).
Casey, K., Glennerster, R., & Miguel, E. (2012). Reshaping institutions: Evidence on aid impactsusing a preanalysis plan. Quarterly Journal of Economics, 127 (4).
Chattopadhyay, R. & Duflo, E. (2004). Women as policy makers: Evidence from a randomizedpolicy experiment in India. Econometrica, 72 (5).
Fearon, J. D., Humphreys, M., & Weinstein, J. M. (2009). Can development aid contribute to socialcohesion after civil war? Evidence from a field experiment in post-conflict Liberia. AmericanEconomic Review: Papers and Proceedings, 99 (2).
Fearon, J. D., Humphreys, M., & Weinstein, J. M. (2011, December). Democratic institutions andcollective action capacity: Results from a field experiment in post-conflict Liberia.
Freedman, D. A. (2008). On regression adjustments to experimental data. Applied Mathematics,40, 180–193.
Fritzen, S. (2007). Can the design of community-driven development reduce the risk of elite capture?Evidence from Indonesia. World Development, 35 (8), 1359–1375.
Fung, A. & Wright, E. O. (2003). Deepening democracy: Institutional innovations in empoweredparticipatory governance. Verso.
Hossain, M. A., Sengupta, M. K., Ahamed, S., Rahman, M. M., Mondal, D., Lodh, D., . . .Chakraborti, D. (2005). Ineffectiveness and poor reliability of arsenic removal plants in WestBengal, India. Environmnetal Science Technology, 39 (11), 4300–4306.
Humphreys, M., de la Sierra, R. S., & van der Windt, P. (2012, June). Social and economic impactsof Tuungane: Final report on the effects of a community driven reconstruction program inEastern Democratic Republic of Congo.
Isham, J., Narayan, D., & Pritchett, L. (1995). Does participation improve performance? Estab-lishing causality with subjective data. World Bank Economic Review, 9 (2), 175–2000.
Khwaja, A. (2004). Is increasing community participation always a good thing? Journal of theEuropean Economic Association: Papers and Proceedings of the Eighteenth Annual Congressof the European Economic Association, 2 (2/3), 427–436.
Khwaja, A. I. (2009). Can good projects succeed in bad communities? Journal of Public Economics,93.
Kleemeier, E. (2000). The impact of participation on sustainability: An analysis of the Malawi ruralpiped scheme program. World Development, 28 (5), 929–944.
27
Mansuri, G. & Rao, V. (2013). Localizing development: Does participation work? Washington, DC:World Bank.
Olken, B. A. (2007). Monitoring corruption: Evidence from a field experiment in Indonesia. Journalof Political Economy, 115 (2).
Olken, B. A. (2010). Direct democracy and local public goods: Evidence from a field experiment inIndonesia. American Political Science Review, 104 (2).
Olson, M. (1971). The logic of collective action. Harvard University Press.Pradhan, M., Suryadarma, D., Beatty, A., Wong, M., Gaduh, A., Alisjahbana, A., & Artha, R. P.
(2014). Improving educational quality through enhancing community participation: Resultsfrom a randomized field experiment in Indonesia. American Economic Journal: Applied Eco-nomics, 6 (2).
Rao, V. & Ibanez, A. M. (2005). The social impact of social funds in Jamaica: A ‘participatoryeconometric’ analysis of targeting, collective action, and participation in community-drivendevelopment. Journal of Development Studies, 41 (5), 788–838.
Sara, J. & Katz, T. (1997). Making rural water sustainable: Report on the impact of project rules.UNDP – World Bank Water and Sanitation Program.
Smith, A., Lingas, E., & Rahman, M. (2000). Contamination of drinking water by arsenic inBangladesh: A public health emergency. Bulletin World Health Organization, 78, 1093–1103.
Stiglitz, J. (2002). Participation and development: Perspectives from the comprehensive develop-ment paradigm. Review of Development Economics, 6 (2), 163–182.
UNICEF. (2008, October). Arsenic mitigation in bangladesh.World Bank. (2004). World Development Report 2004: Making services work for poor people. Wash-
ington, DC: The World Bank.
28
Figures and Tables
Figure 1: Heterogeneity in average treatment effect with village size
-.20
.2.4
Cha
nge
in re
porte
d ac
cess
to s
afe
drin
king
wat
er
0 500 1000Number of households in village
Treated Control
Graph shows results from a local linear regression of the change in access to safe drinking water on villagesize for treated and control villages. 90% confidence intervals are cluster bootstapped at the village level.
29
Figure 2: Heterogeneity in treatment effect by decision-making model with village size
-.20
.2.4
Chan
ge in
repo
rted
acce
ss to
saf
e dr
inkin
g wa
ter
0 500 1000Number of households in village
NGO CP TD
Graph shows results from a local linear regression of the change in access to safe drinking water on villagesize for treated and control villages.
30
Table 1: Treated vs ControlBaseline Summary Statistics and Randomization Checks
Control Treated Treatment - Control(1) (2) (3) (4) (5)
Proportion of villages inGopalganj
Mean 0.51 0.55s.e. (0.05) (0.04)
Proportion of villages in SouthMatlab
Mean 0.00 0.23***s.e. - (0.04)
No of households in villageMean 245 221 -26 -29 -34
s.e. (21) (17) (29) (27) (31)
% of water sources arseniccontaminated
Mean 0.96 0.95 0.01 0.00 0.00s.e. (0.01) (0.01) (0.01) (0.01) (0.01)
Reports using arsenic safe waterMean 0.55 0.41*** -0.01 -0.03 -0.03
s.e. (0.04) (0.03) (0.04) (0.03) (0.04)
Changed source of drinking waterdue to arsenic in last 5 years?
Mean 0.49 0.35*** 0.00 -0.02 -0.03s.e. (0.04) (0.03) (0.03) (0.03) (0.04)
Anyone in household hassymptoms of arsenic poisoning?
Mean 0.009 0.009 -0.001 0.000 0.000s.e. (0.002) (0.001) (0.003) (0.002) (0.003)
Total value of household assetsMean 570284 540165 -14573 -5071 -5865
s.e. (30362) (21720) (41233) (37175) (42881)
Access to electricity?Mean 0.46 0.39 -0.05 -0.02 -0.03
s.e. (0.03) (0.03) (0.05) (0.04) (0.05)
Household head literateMean 0.61 0.60 0.01 0.00 0.00
s.e. (0.02) (0.02) (0.03) (0.03) (0.03)
Household head MuslimMean 0.70 0.70 0.04 0.05 0.05
s.e. (0.04) (0.04) (0.05) (0.05) (0.06)
Household head farmerMean 0.42 0.45 0.03 0.02 0.03
s.e. (0.02) (0.01) (0.02) (0.02) (0.02)
Number of associations incommunity
Mean 6.25 6.30 -0.18 -0.18 -0.20s.e. (0.14) (0.15) (0.22) (0.19) (0.22)
Number of collective actions incommunity
Mean 0.89 0.96 0.05 0.07 0.08s.e. (0.08) (0.09) (0.05) (0.06) (0.06)
Number of villages 99 127 197 226 226Number of households 3914 4976 7755 8890 8890
Note: P-values test significance of differences between treated and control villages, controlling for the dif-ferent treatment proportions in Gopalganj, North and South Matlab in columns 3-5). Data in rows 1) and2) come from project records. Data in rows 3) and 4) comes from data from the Bangladesh Arsenic Miti-gation Water Supply Project. All other data is from baseline household surveys. Two villages are missingall baseline data. Standard errors (in parentheses) are robust, and clustered at the village level in rows 5)onwards.*** p<0.01, ** p<0.05, * p<0.1.
31
Table 2: Decision-making structures
Non-participatory Top Down (TD) Project staff took all project decisions, after an extended (typi-cally 2-day) period of information gathering, using the followingcriteria to decide water source location:
• public/convenient location• population density• existing safe water options
Participatory CommunityParticipation (CP)
The community took all project decisions using their own (unob-served) decision-making structures, following a community-wideinformation meeting led by project staff.
NGO-FacilitatedCommunity
Participation (CP)
The community took all project decisions at a community- widemeeting, following smaller information meetings for differentgroups. We imposed two decision-making rules. If decisions madedid not satisfy these rules, project staff did not implement the de-cisions:
• Attendance at the community meeting had to include: atleast 10 men, of which 5 had to qualify as poor; and atleast 10 women, of which 5 had to qualify as poor.
• Decisions had to be unanimous.
Table 3: Technologies to provide arsenic-safe drinking water
Required community contribution persafe water source installed
Technology Cost 1 2 3
Deep tubewell (DTW) 50000 4500 6000 7500
Shallow tubewell (STW) 20000 3000 3500 4000
Arsenic-Iron Removal Plant (AIRP) 60000 6000 7500 N/A
Deep-set tubewell (DSTW) 60000 6000 7500 N/A
Note: All prices in Bangladeshi Taka. 1 US$≈ 80BDT.
32
Table 4: Assignment to decision-making structureBaseline Summary Statistics and Randomization Checks
NGO CP TD(1) (2) (3)
Proportion of villages in GopalganjMean 0.55 0.55 0.56
s.e. (0.08) (0.08) (0.08)
Proportion of villages in South MatlabMean 0.24 0.21 0.23
s.e. (0.07) (0.06) (0.06)
No of households in villageMean 213 213 236
s.e. (33) (24) (32)
% of water sources arseniccontaminated
Mean 0.96 0.95 0.95s.e. (0.01) (0.01) (0.01)
AIRPs only feasible technologyMean 0.10 0.14 0.14
s.e. (0.05) (0.05) (0.05)
Reports using arsenic safe waterMean 0.36 0.42 0.44
s.e. (0.05) (0.05) (0.05)
Changed source of drinking water dueto arsenic in last 5 years?
Mean 0.32 0.35 0.37s.e. (0.05) (0.05) (0.05)
Anyone in household has symptoms ofarsenic poisoning?
Mean 0.004** 0.009 0.012*s.e. (0.002) (0.003) (0.003)
Total value of household assetsMean 543364 548174 529180
s.e. (39597) (42208) (30413)
Access to electricity?Mean 0.37 0.39 0.42
s.e. (0.05) (0.05) (0.05)
Household head literateMean 0.60 0.58 0.62
s.e. (0.03) (0.03) (0.02)
Household head MuslimMean 0.68 0.70 0.73
s.e. (0.07) (0.06) (0.06)
Household head farmerMean 0.44 0.46 0.44
s.e. (0.03) (0.02) (0.02)
Number of associations in communityMean 6.36 6.05 6.47
s.e. (0.25) (0.18) (0.31)
Number of collective actions incommunity
Mean 0.91 1.00 0.98s.e. (0.14) (0.16) (0.15)
Number of villages 42 42 43Number of households 1638 1635 1703
Note: P-values test significance of the difference between model and other treated villages. Data fromhousehold surveys except rows 1), 2) and 5) which come from project records and rows 3) and 4) whichcome from the Bangladesh Arsenic Mitigation Water Supply Project. Baseline data for one CP village ismissing. Standard errors (in parentheses) are robust, and clustered at the village level in rows 5) onwards.*** p<0.01, ** p<0.05, * p<0.1.
33
Table 5: Participation in project decision-making
Fraction of meeting participants:No. of
participantsFemale
Low s.e.status
No 2◦
education
(1) (2) (3) (4)
All treated Mean 30.7 0.27 0.42 0.60s.e. (1.1) (0.02) (0.02) (0.02)
NGO Mean 33.6 0.31 0.44 0.64s.e. (2.5) (0.03) (0.04) (0.03)
CP Mean 29.9 0.29 0.46 0.55s.e. (1.4) (0.03) (0.02) (0.03)
TD Mean 28.5 0.21 0.36 0.61s.e. (1.8) (0.03) (0.03) (0.03)
NGO = CP p-value 0.200 0.615 0.570 0.031CP = TD p-value 0.561 0.073 0.006 0.154TD = NGO p-value 0.111 0.023 0.085 0.490
N 126 123 119 122
Note: P-values test pairwise significance of the difference between the means acrossmodels indicated, from a regression of the outcome variable on indicators for the threetypes of treatment (with no constant). Robust standard errors shown in parentheses.In two villages, meetings were not held.
34
Table 6: Project Outcomes
Outcome Variable
Water sourcesinstalled
Installed inpublic places
Number ofcontributors
(1) (2) (3)
Panel A: All villages
All treated Mean 2.14 1.38 7.81s.e. (0.10) (0.10) (0.83)
NGO Mean 2.21 1.19 9.37s.e. (0.16) (0.14) (1.62)
CP Mean 2.21 1.00 5.40s.e. (0.17) (0.14) (0.93)
TD Mean 2.00 1.93 8.67s.e. (0.18) (0.18) (1.63)
NGO = CP p-value 1.000 0.334 0.036CP = TD p-value 0.394 0.000 0.084TD = NGO p-value 0.376 0.002 0.764
N 127 127 126
Panel B: Villages where tubewells feasible
All treated Mean 2.45 1.58 9.02s.e. (0.08) (0.10) (0.92)
NGO Mean 2.46 1.30 10.61s.e. (0.13) (0.14) (1.74)
CP Mean 2.53 1.14 6.11s.e. (0.14) (0.15) (1.03)
TD Mean 2.36 2.31 10.33s.e. (0.16) (0.15) (1.82)
NGO = CP p-value 0.718 0.440 0.028CP = TD p-value 0.428 0.000 0.046TD = NGO p-value 0.624 0.000 0.912
N 109 109 108
Note: P-values test pairwise significance of the difference between the means across modelsindicated, from a regression of the outcome variable on indicators for the three types oftreatment (with no constant). Robust tandard errors shown in parentheses. In villageswhere no water sources were installed, the number of contributors is coded as zero. In onevillage, the number of contributors was not recorded.
35
Table 7: Estimates of average treatment effect
OLS OLS IV OLS OLS IV OLS(1) (2) (3) (4) (5) (6) (7)
Panel A: Reported access to safe drinking water at baseline
Treated Coefficient -0.141*** -0.011 -0.031 -0.181*** -0.017 -0.042 -0.024s.e. (0.05) (0.04) (0.04) (0.05) (0.04) (0.04) (0.10)
Control Coefficient 0.547 0.830 0.837 0.609 0.833 0.842 0.253s.e. (0.04) (0.03) (0.03) (0.04) (0.03) (0.03) (0.07)
First-stage F-test 1340 896N 8695 7608 8695 7375 6288 7375 1242
Panel B: Reported access to safe drinking water at follow-up
Treated Coefficient 0.022 0.141*** 0.132*** 0.001 0.148*** 0.137*** -0.007s.e. (0.05) (0.03) (0.04) (0.05) (0.04) (0.04) (0.08)
Control Coefficient 0.533 0.820 0.823 0.607 0.817 0.821 0.201s.e. (0.04) (0.02) (0.03) (0.04) (0.03) (0.03) (0.06)
First-stage F-test 1335 896N 8442 7387 8442 7168 6113 7168 1201
Panel C: Change in reported access to safe drinking water
Treated Coefficient 0.164*** 0.148*** 0.157*** 0.182*** 0.164*** 0.175*** 0.004s.e. (0.03) (0.03) (0.04) (0.03) (0.04) (0.04) (0.09)
Control Coefficient -0.014 -0.012 -0.015 -0.003 -0.017 -0.021 -0.042s.e. (0.02) (0.02) (0.02) (0.02) (0.02) (0.02) (0.05)
First-stage F-test 1337 897N 8427 7375 8427 7154 6102 7154 1200
Feasible technology All All All Tubewell Tubewell Tubewell AIRPControls for upazila No Yes Yes No Yes Yes -Includes S. Matlab Yes No Yes Yes No Yes -
Control villages All All All Matched Matched Matched Matched
Note: Treatment is instrumented using synthetic assignment to treatment in South Matlab in columns 3)and 6). In columns 4) to 7) the control group is matched to the subset of treated villages using baselinepropensity score matching. Survey weights are applied so that each village counts equally in the analysis.Controls for upazila include an indicator for Gopalganj, and an indicator for South Matlab, where included.Standard errors (in parentheses) are robust and clustered at the village level.*** p<0.01, ** p<0.05, * p<0.1.
36
Table 8: Estimates of treatment effect by decision-making model
OLS OLS IV OLS OLS IV OLS(1) (2) (3) (4) (5) (6) (7)
Panel A: Reported access to safe drinking water at baseline
NGO Coefficient -0.19*** -0.06 -0.07 -0.24*** -0.08 -0.09 -0.10s.e. (0.06) (0.05) (0.05) (0.07) (0.06) (0.06) (0.09)
CP Coefficient -0.13** 0.01 -0.02 -0.17** 0.00 -0.04 0.04s.e. (0.07) (0.06) (0.05) (0.07) (0.06) (0.06) (0.17)
TD Coefficient -0.10 0.02 0.00 -0.13* 0.03 0.01 -0.04s.e. (0.06) (0.05) (0.05) (0.07) (0.05) (0.06) (0.10)
Constant Coefficient 0.55 0.83 0.84 0.61 0.83 0.84 0.25s.e. (0.04) (0.03) (0.03) (0.04) (0.03) (0.03) (0.07)
NGO = CP 0.437 0.295 0.460 0.395 0.257 0.443 0.389CP = TD 0.708 0.916 0.673 0.563 0.683 0.456 0.609
TD = NGO 0.239 0.199 0.208 0.151 0.106 0.119 0.540
NGO = pooled 0.258 0.180 0.252 0.185 0.114 0.180 0.327CP = pooled 0.819 0.587 0.851 0.878 0.673 0.988 0.480TD = pooled 0.372 0.423 0.329 0.246 0.241 0.183 0.873
N 8695 7608 8695 7375 6288 7375 1242
Feasible technology All All All Tubewell Tubewell Tubewell AIRPControls for upazila No Yes Yes No Yes Yes -Includes S. Matlab Yes No Yes Yes No Yes -
Control villages All All All Matched Matched Matched Matched
Note: Model is instrumented using model interacted with synthetic assignment to treatment in South Matlabin columns 3) and 6). In columns 4) to 7) the control group is matched to the subset of treated villagesusing baseline propensity score matching. Survey weights are applied so that each village counts equally inthe analysis. Controls for upazila include an indicator for Gopalganj, and an indicator for South Matlab,where included. Standard errors (in parentheses) are robust and clustered at the village level.*** p<0.01, ** p<0.05, * p<0.1.
37
Table 8, continued: Estimates of treatment effect by decision-making model
OLS OLS IV OLS OLS IV OLS(1) (2) (3) (4) (5) (6) (7)
Panel B: Reported access to safe drinking water at follow-up
NGO Coefficient 0.03 0.15*** 0.16*** 0.01 0.16*** 0.16*** -0.13*s.e. (0.06) (0.06) (0.06) (0.06) (0.06) (0.06) (0.07)
CP Coefficient 0.00 0.14*** 0.11** -0.03 0.14** 0.10* 0.09s.e. (0.06) (0.05) (0.05) (0.07) (0.05) (0.05) (0.13)
TD Coefficient 0.03 0.13*** 0.13*** 0.02 0.15*** 0.15*** -0.02s.e. (0.06) (0.05) (0.05) (0.06) (0.05) (0.05) (0.09)
Constant Coefficient 0.53 0.82 0.82 0.61 0.82 0.82 0.20s.e. (0.04) (0.02) (0.03) (0.04) (0.03) (0.03) (0.06)
NGO = CP 0.702 0.840 0.430 0.657 0.725 0.308 0.091CP = TD 0.697 0.860 0.634 0.499 0.875 0.372 0.473
TD = NGO 0.998 0.708 0.699 0.809 0.823 0.811 0.178
NGO = pooled 0.823 0.748 0.504 0.893 0.745 0.467 0.042CP = pooled 0.660 0.978 0.461 0.528 0.764 0.272 0.217TD = pooled 0.817 0.740 0.968 0.584 0.963 0.704 0.958
N 8442 7387 8442 7168 6113 7168 1201
Feasible technology All All All Tubewell Tubewell Tubewell AIRPControls for upazila No Yes Yes No Yes Yes -Includes S. Matlab Yes No Yes Yes No Yes -
Control villages All All All Matched Matched Matched Matched
Note: Model is instrumented using model interacted with synthetic assignment to treatment in South Matlabin columns 3) and 6). In columns 4) to 7) the control group is matched to the subset of treated villagesusing baseline propensity score matching. Survey weights are applied so that each village counts equally inthe analysis. Controls for upazila include an indicator for Gopalganj, and an indicator for South Matlab,where included. Standard errors (in parentheses) are robust and clustered at the village level.*** p<0.01, ** p<0.05, * p<0.1.
38
Table 8, continued: Estimates of treatment effect by decision-making model
OLS OLS IV OLS OLS IV OLS(1) (2) (3) (4) (5) (6) (7)
Panel C: Change in reported access to safe drinking water
NGO Coefficient 0.22*** 0.21*** 0.22*** 0.25*** 0.24*** 0.25*** -0.05s.e. (0.05) (0.06) (0.06) (0.05) (0.06) (0.06) (0.09)
CP Coefficient 0.14*** 0.13*** 0.13*** 0.15*** 0.13*** 0.14*** 0.03s.e. (0.04) (0.05) (0.05) (0.04) (0.05) (0.05) (0.15)
TD Coefficient 0.13*** 0.11** 0.13*** 0.15*** 0.12** 0.14*** 0.02s.e. (0.04) (0.05) (0.05) (0.04) (0.05) (0.05) (0.12)
Constant Coefficient -0.01 -0.01 -0.02 0.00 -0.02 -0.02 -0.04s.e. (0.02) (0.02) (0.02) (0.02) (0.02) (0.02) (0.05)
NGO = CP 0.178 0.238 0.149 0.116 0.143 0.083 0.653CP = TD 0.946 0.783 0.969 0.954 0.765 0.930 0.947
TD = NGO 0.151 0.160 0.155 0.109 0.104 0.106 0.636
NGO = pooled 0.119 0.150 0.110 0.078 0.088 0.064 0.568CP = pooled 0.429 0.556 0.358 0.338 0.432 0.250 0.786TD = pooled 0.356 0.302 0.388 0.298 0.231 0.329 0.850
N 8427 7375 8427 7154 6102 7154 1200
Feasible technology All All All Tubewell Tubewell Tubewell AIRPControls for upazila No Yes Yes No Yes Yes -Includes S. Matlab Yes No Yes Yes No Yes -
Control villages All All All Matched Matched Matched Matched
Note: Model is instrumented using model interacted with synthetic assignment to treatment in South Matlabin columns 3) and 6). In columns 4) to 7) the control group is matched to the subset of treated villagesusing baseline propensity score matching. Survey weights are applied so that each village counts equally inthe analysis. Controls for upazila include an indicator for Gopalganj, and an indicator for South Matlab,where included. Standard errors (in parentheses) are robust and clustered at the village level.*** p<0.01, ** p<0.05, * p<0.1.
39
Tab
le9:
Rob
ust
nes
sch
ecks
OL
SO
LS
IVIV
IVO
LS
IVIV
IVIV
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
NG
OC
oeffi
cien
t0.2
4**
*0.
37***
0.10
0.2
5***
0.2
7***
0.2
4***
0.4
6***
0.2
8***
0.2
8***
0.3
0***
s.e.
(0.0
5)
(0.0
9)(0
.07)
(0.0
6)
(0.0
6)
(0.0
6)
(0.1
3)(0
.07)
(0.0
7)(0
.07)
CP
Coeffi
cien
t0.1
5**
*0.
19*
**
0.0
80.
14*
**
0.15*
**
0.14*
**
0.1
80.1
5***
0.1
6**
0.1
3***
s.e.
(0.0
4)
(0.0
7)(0
.07)
(0.0
5)
(0.0
5)
(0.0
4)
(0.1
6)(0
.06)
(0.0
7)(0
.05)
TD
Coeffi
cien
t0.1
4**
*0.2
1**
*0.0
60.1
4**
*0.1
6**
*0.1
2**
0.2
1*
0.1
6***
0.2
0***
0.1
4***
s.e.
(0.0
4)
(0.0
7)(0
.06)
(0.0
5)
(0.0
5)
(0.0
5)
(0.1
2)(0
.06)
(0.0
7)(0
.05)
Con
stant
Coeffi
cien
t-0
.04
0.0
1-0
.02
-0.0
3-0
.03
-0.0
30.2
00.0
2s.
e.(0
.04)
(0.0
2)
(0.0
2)
(0.0
2)
(0.0
6)
(0.0
2)
(0.0
4)
(0.0
3)
NG
O=
CP
0.1
180.
055
0.7
84
0.0
83
0.084
0.1
65
0.1
33
0.0
98
0.1
17
0.0
20
NG
O=
TD
0.1
110.
107
0.6
37
0.1
06
0.108
0.1
22
0.1
17
0.1
51
0.2
80
0.0
27
NG
O=
poole
d0.0
800.
052
0.6
74
0.0
64
0.065
0.1
06
0.0
77
0.0
86
0.1
30
0.0
13
N401
7320
139
53
734
277
84
466
08347
8357
7154
7154
Mea
sure
ofac
cess
Rep
ort
edR
eport
edR
eport
edR
eport
edR
eport
edR
eport
edR
eport
edR
eport
edV
erifi
edC
om
bin
edF
easi
ble
tech
nolo
gyT
ub
ewel
lT
ub
ewel
lT
ub
ewel
lT
ub
ewel
lT
ub
ewel
lA
llA
llA
llT
ub
ewel
lT
ub
ewel
lSam
ple
All
Gop
alg
anj
Matl
abA
llA
llA
llA
llA
llA
llA
llU
pazi
laC
ontr
ols
No
-Y
esY
esY
esN
oY
esY
esY
esY
esA
ddit
ional
contr
ols
No
No
No
No
No
Tec
hnolo
gy
Vil
l.si
zeA
sset
sN
oN
oC
ontr
ol
villa
ges
None
Matc
hed
All
Alt
ernat
ive
All
None
All
All
Matc
hed
Matc
hed
Note
:M
od
elis
inst
rum
ente
du
sin
gm
od
elin
tera
cted
wit
hsy
nth
etic
ass
ign
men
tto
trea
tmen
tin
Sou
thM
atl
ab
wh
ere
ind
icate
d.
Su
rvey
wei
ghts
are
app
lied
soth
atea
chvil
lage
cou
nts
equ
ally
inth
ean
aly
sis.
Contr
ols
for
up
azi
lain
clu
de
an
ind
icato
rfo
rG
op
alg
anj,
an
dan
ind
icato
rfo
rS
ou
thM
atl
ab
,w
her
ein
clu
ded
.W
her
ead
dit
ion
alco
ntr
ols
are
incl
ud
ed,
the
inte
ract
ion
sb
etw
een
the
contr
ols
an
dth
etr
eatm
ent
ind
icato
rsare
als
oS
tan
dar
der
rors
(in
par
enth
eses
)ar
ero
bu
stan
dcl
ust
ered
at
the
vil
lage
leve
l.**
*p<
0.01
,**
p<
0.05
,*
p<
0.1.
40
Appendices
Appendix A: Data
Variable Construction
We asked the households to list all the sources of water they used for drinking and cooking.In the analysis, we focus on the most important source of water for drinking and cooking, whichwe asked households to list first. We also asked households to report the percentage of water fordrinking and cooking that they obtained from each source, but results based on the source fromwhich households report drawing the largest percentage of water are unstable between baseline andfollowup, whereas the results based on the first-listed water source are more consistent. This maybe attributable to slight differences in the way in which the question was asked as to whether thequestion referred to water used for drinking only or drinking and cooking.
Reports using safe drinking water If the household reports using a tubewell, we code thehousehold as reporting using safe water if they report that the source is arsenic-safe, and reportingunsafe water if it is unsafe or if they dont know the source’s safety. If the household reports usingan unsafe source with respect to bacterial contamination (i.e. a dug well or surface water), we codethe household as reporting using unsafe water. Some sources can be presumed to be safe from bothbacterial and arsenic contamination (e.g. AIRPs, PSF, rainwater, deep-set tubewells). In thesecases, we code the household as reporting using safe water unless the household reports that thewater is unsafe. The numbers of households using these sources is small. If the household reportsusing any other source, we code the household as reporting using safe water if they report that thesource is safe, and reporting unsafe water if it is unsafe or if they dont know the source’s safetystatus.
Reports using verified safe drinking water Many tubewells in Bangladesh have been testedfor arsenic safety in the past and marked with green (safe) or red (unsafe). If households reportedusing a tubewell, enumerators visited the tubewell to confirm whether it was marked safe, unsafe,or not marked, as long as the tubewell was less than 5 minutes walk away. However, an earlyversion of the baseline survey used did not include this question, so this information is missing forsome villages at baseline. We code this variable as follows. We always code households reportingan unsafe source as not using a verified source of safe drinking water. If the household reportsusing a source that can otherwise be presumed to be safe (i.e. rainwater), we code the householdas using a verified source of safe drinking water, unless they report it to be unsafe. If they reportusing a tubewell, we code the house as using a verified source of safe drinking water if the source isverifiably safe, and not using a verified source of safe drinking water if the tubewell is either markedunsafe or can’t be verified safe (either because it is too far, or because it is unmarked, or becausethe question wasn’t asked in this village at baseline). If the household reports using any othersource, we code the household as reporting using a verified source safe water if the enumeratorsrecorded that the source could be verified safe, and reporting using an unverified or unsafe sourceif the source is unsafe or if it cannot be verified.
Combined measure We combine these measures by using the verified data, when the safetyof the source could be verified, and the reported data, when the safety of the source could not beverified.
Construction of matched control groups in Gopalganj
Tubewells are not feasible where there is a rocky layer separating the surface from the arsenic-safe deep aquifer, which cannot be penetrated with local drilling technologies. This was only a
41
problem in this study in Gopalgnaj, as tubewells of varying kinds were feasible in all villages inMatlab. The following discussion is therefore limited to Gopalganj. There is substantial spatialcorrelation in the location of these rocky layers.
In Gopalganj, there was no overall problem with random assignment to treatment. AppendixTable B1, columns 1) and 2) confirm this; of 12 tests comparing treated to control villages in Gopal-ganj, only one shows statistically significant differences at the 10% level, which is approximatelywhat we would expect due to chance. However, when we compare either the AIRP villages (column3) or the tubewell villages (column 6), to the full group of control villages, there is some evidencethat feasible technology is correlated with other village level characteristics. In column 3), only 1of 12 tests shows statistically significant differences at the 10% level, but in column 6), 1 test showsstatistically significant differences at the 5% level, and additionally another shows statistically sig-nificant differences at the 10% level. To be conservatve, we construct matched control groups forthe AIRP villages and tubewell villages, primarily exploiting the spatial correlation in location ofthe rocky layer.
In Gopalganj, there are 18 unions (the smallest rural administrative and local government unitsin Bangladesh). In three of these unions, only AIRPs were feasible in all treated villages. In 4unions, tubewells were feasible in all treated villages. First, we assign the five control villages inunions where only AIRPs were feasible to the AIRP-matched control group. Second, we assign the18 control villages in unions where tubewells were always feasible to the tubewell-matched controlgroup. For the remainder of the unions — for which tubewells were feasible in a fraction of thevillages — we use two strategies to identify matched control villages.
First, we construct a propensity score index by running a logit regression of an indicator forwhether tubewells (or only AIRPs) were feasible on a set of baseline characteristics which we observefor both treated and control groups and which are correlated feasible technology in the treatmentgroup. These baseline characteristics are: union fixed effects, access to electricity, number oftubewells operative, number of tubewells arsenic contaminated, reported access to safe drinkingwater at baseline, verifiability of arsenic safety of primary reported source, verified access to safedrinking water at baseline, fraction of village changing source of safe drinking water in the precedingfive years due to arsenic contamination; whether any member of the household has arsenic poisoning;mean literacy of household heads; and number of coordinated actions in the village. We then assignthe remaining villages to the relevant matched control group where their propensity score is greaterthan 0.5 (although in reality the propensity scores are strongly clustered around 0 and 1). Thisprocess assigns a total of 16 villages to the AIRP-matched control group, and 33 villages to thetubewell-matched control group. 3 villages are not assigned to either control group, and 2 villagesare assigned to both groups. Columns 4) and 7) show the characteristics of the groups constructedin this way. No tests show statistically significant differences between these groups and the AIRPor tubewell villages from the treated groups.
The second method we use is to assign the villages in each union to the tubewell-matchedand AIRP-matched control groups according to the observed, union-level probability of AIRPand tubewell feasibility. We do this by generating a random number for each control village andassigning them to the relevant control group if the random number is less than the proportion oftreated villages in that union in which the technology was feasible. This process assigns 9 villagesto the AIRP-matched control villages and 40 villages to the tubewell-matched control villages.One village is not assigned to either control group, and by construction none are assigned to bothgroups. Columns 5) and 8) show the characteristics of the groups constructed in this way. 3 testsshow statistically significant differences between the AIRP-matched control villages and the AIRPvillages from the treated groups, but no tests show statistically significant differences between thetubewell-matched control villages constructed in this way and the tubewell villages from the treated
42
groups.For the main specification, we therefore use the matched control group created using propensity
score matching, but we will compare these results to results using the full control group and thesecond matching strategy in robustness checks.
43
Appendix B: Figures and Tables
44
Tab
leB
1:
Base
line
Su
mm
ary
Sta
tist
ics
for
AIR
Pan
dn
on-A
IRP
villa
ges
inG
opal
gan
j
Tre
ated
Contr
ol
AIR
PA
IRP
Contr
ol
1A
IRP
Contr
ol
2T
ub
ewel
lT
ub
ewel
lC
ontr
ol1
Tub
ewel
lC
ontr
ol
2(1
)(2
)(3
)(4
)(5
)(6
)(7
)(8
)
No
of
house
hol
ds
invilla
ge26
2257
338
268
271
242
255
263
(26)
(33)
(71)
(44)
(57)
(27)
(45)
(41)
%of
wat
erso
urc
esar
senic
conta
min
ated
0.97
0.9
60.9
70.9
70.9
9*
0.96
0.9
60.9
6(0
.01)
(0.0
1)
(0.0
1)
(0.0
1)
(0.0
1)
(0.0
1)(0
.01)
(0.0
1)
Rep
orts
usi
ng
arse
nic
safe
wat
er0.
270.2
60.2
30.2
50.2
60.
27
0.2
70.2
6(0
.03)
(0.0
4)
(0.0
7)
(0.0
7)
(0.1
1)
(0.0
4)(0
.04)
(0.0
4)
Chan
ged
sourc
eof
dri
nkin
gw
ater
due
toar
senic
inla
st5
year
s?0.
210.2
00.2
20.2
10.1
9*
0.19
0.1
80.1
9(0
.03)
(0.0
3)
(0.0
6)
(0.0
7)
(0.1
0)
(0.0
3)(0
.04)
(0.0
4)
Anyo
ne
inhou
sehol
dhas
sym
pto
ms
ofar
senic
poi
sonin
g?0.
006
0.0
09
0.0
05
0.0
08
0.0
06
0.006
0.0
030.0
08
(0.0
02)
(0.0
03)
(0.0
02)
(0.0
04)
(0.0
04)
(0.0
02)
(0.0
02)
(0.0
03)
Tot
alva
lue
ofhou
sehol
das
sets
4692
30512028
568106
531573
527437
436
844
4930
45
5145
95(2
1450
)(4
0937)
(47317)
(82727)
(146437)
(232
79)
(471
48)
(412
89)
Acc
ess
toel
ectr
icit
y?
0.37
0.4
50.6
0*
0.6
80.6
30.
31*
*0.3
50.
41(0
.04)
(0.0
4)
(0.0
7)
(0.0
3)
(0.0
5)
(0.0
4)
(0.0
5)
(0.0
5)
Hou
sehol
dhea
dlite
rate
0.58
0.5
60.5
40.4
90.4
50.5
90.6
20.
58(0
.03)
(0.0
4)
(0.0
6)
(0.0
7)
(0.0
8)
(0.0
3)
(0.0
4)
(0.0
4)
Hou
sehol
dhea
dM
usl
im0.
550.4
80.6
60.4
60.5
20.5
20.4
60.
48(0
.05)
(0.0
6)
(0.1
1)
(0.1
0)
(0.1
4)
(0.0
6)
(0.0
8)
(0.0
7)
Hou
sehol
dhea
dfa
rmer
0.50
0.4
60.4
50.3
90.3
90.
51*
0.4
90.
47(0
.02)
(0.0
2)
(0.0
4)
(0.0
3)
(0.0
5)
(0.0
2)
(0.0
3)
(0.0
2)
Num
ber
ofas
soci
atio
ns
inco
mm
unit
y6.
766.8
66.6
57.1
87.6
26.7
76.7
96.
82(0
.23)
(0.2
2)
(0.3
6)
(0.5
0)
(0.7
0)
(0.2
9)
(0.2
3)
(0.2
2)
Num
ber
ofco
llec
tive
acti
ons
inco
mm
unit
y0.
230.1
4*
0.2
60.1
20.1
0*
0.2
20.1
50.
14(0
.04)
(0.0
2)
(0.0
8)
(0.0
4)
(0.0
4)
(0.0
5)
(0.0
3)
(0.0
3)
Num
ber
ofvilla
ges
7050
16
16
952
3338
Num
ber
ofhou
sehol
ds
3859
196
91102
627
347
267
713
0415
04
Note
:S
tars
ind
icat
esi
gnifi
can
ceof
test
sof
diff
eren
ceof
mea
ns:
bet
wee
ntr
eate
dan
dco
ntr
ol
vil
lages
inG
op
alg
an
jin
colu
mn
2);
bet
wee
nA
IRP
an
dal
lco
ntr
olvil
lage
sin
colu
mn
3);
bet
wee
nA
IRP
an
dm
atc
hed
contr
ol
vil
lages
inco
lum
ns
4)
an
d5);
bet
wee
ntu
bew
ell
an
dall
contr
ol
vil
lages
inco
lum
n6;
and
bet
wee
ntu
bew
ell
and
mat
ched
contr
olvil
lages
inco
lum
ns
7)
an
d8).
Data
inro
ws
1)
an
d2)
com
efr
om
the
Ban
gla
des
hA
rsen
icM
itig
ati
on
Wat
erSu
pp
lyP
roje
ct.
Th
ere
mai
nin
gd
ata
are
from
hou
seh
old
surv
eys.
Sta
nd
ard
erro
rs(s
how
nin
pare
nth
eses
)are
clu
ster
edat
the
vil
lage
leve
l.***
p<
0.01
,**
p<
0.05
,*
p<
0.1.
45